Grok Bot Routines: Which AI Models To Pair With Them, With Real GoldieBench Scores

Grok Bot routines don't let you choose the AI model, so the GoldieBench scores that matter are for the models you use around them: to draft the brief, to check the output, or to run a copy of the job somewhere you do pick the model.

xAI's settings page says it plainly: Grok Bot manages model selection, so there is no model picker.

That changes the question from "which model runs my routine" to "which model do I pair with it".

Every score below is quoted from a live GoldieBench page I read on 31 October 2026, with the page linked.

If you want my team to help you plan the setup, book a free SEO strategy session.

What a routine is, in one paragraph

A routine tells one Grok Bot when to run a saved workflow; per the skills and routines docs, it fires on a schedule or on a Slack, email, webhook, GitHub, Linear, Sentry or PagerDuty event.

Each Bot can own up to 50, scheduled runs need at least five minutes between them, and the app holds the 20 newest run records for each one.

Since v0.75.5 on 23 October 2026, the changelog allows 300 runs an hour across all your routines and 200 for any single routine.

And every run draws on your plan's usage.

What GoldieBench measures, and what it leaves out

The methodology page describes each task as one fixed prompt, one shot, one self-contained HTML file, scored 0 to 10.

The Grok 4.7 deep dive is direct about the limits: the bench does not measure multi-turn agent work, long-context recall or API latency.

Routines are multi-turn agent work.

So read these scores as a signal of how well a model writes and builds from a single clear instruction, which is exactly what a routine brief is.

Don't read them as a measure of how a model behaves across a long unattended job.

The scores worth knowing for routine work

The models index also lists Claude Opus 5 at 8.27 and GPT-5.6 Sol at 8.16.

One note on Grok: the models index card shows 47 tasks, while the Grok benchmark page counts 43 scored tasks, so I quote the 43 with the 8.09 average from that page.

Grok and Grok 4.7: same family, very different boards

Grok scores 8.09 across 43 tasks on its benchmark page, with 5 golds, 1 silver and 1 bronze.

Its category rows read Games 8.3, Pages 8.3, Sims 8.0 and Visuals 7.7.

Grok 4.7 sits at 7.15 across 20 tasks, all of them games, with 0 golds, 2 silvers and 3 bronzes.

Its weakest builds were Voxelcraft at 2.2 and Crypt at 2.3, both scenes that never rendered.

If you want a Grok model to draft routine briefs or review Bot output outside Grok Bot, the stronger record belongs to Grok.

Hermes MoA and Fusion: panels sit at the top

Fusion leads the whole board at 8.59, with 21 task golds out of 47 scored tasks.

Hermes MoA averages 8.17 and ranks third, with category rows of Games 8.0, Pages 8.0, Sims 8.4 and Visuals 8.4.

Both are multi-model panels rather than single models, according to their cards on the models index.

On head-to-head records, the Grok page shows Grok trailing Fusion 2-31 with 10 ties, and trailing Hermes MoA 14-27 with 2 ties.

For one-shot quality, a panel is the safest pick for writing the skill your routine will run.

Kimi K3 for long briefs and big inputs

Kimi K3 scores 7.89 over 50 scored tasks and ranks 11th.

Its benchmark page lists a context of 1,048,576 tokens, and its models index card calls it tuned for long-horizon agent work.

That line is the card description, not a result the one-shot bench measured.

Where it helps is reading a long run history, a big inbox export or a full skill library in one pass while you review routines.

Where you really choose the model: Hermes cron jobs

If you run a copy of a routine as a Hermes scheduled task, the model is your call.

The Hermes docs set the order a job resolves its model at fire time: a per-job pin first, then cron.model, then your main agent model.

Pin a job with hermes cron edit <job_id> --pin, or set a fleet default with hermes config set cron.model <name>.

A job can also pin its own reasoning effort, so a heavy weekly analysis can run high while a cheap hourly check runs minimal.

That is where these GoldieBench averages turn into a real decision.

A pairing plan you can copy

1. Write the skill with a top-board model such as Fusion or Hermes MoA, then save it in Grok Bot.
2. Run the routine in Grok Bot, which picks its own model.
3. Review the results weekly with a long-context model such as Kimi K3, pasting in the run history.
4. Mirror critical jobs as pinned Hermes cron jobs on the model you trust most.

Keep every send, purchase and deletion behind approval on both sides.

The approvals docs say routine email waits as a draft until you press Send.

Limits beat model choice on a busy routine

A better model can't save a routine that fires too often.

The routines docs say each run spends usage, and a short interval or a busy Slack channel can drain a week's worth in one day.

A run past an hourly cap fails with Automation run rate limit exceeded and needs a manual rerun.

Fix the trigger first, then worry about the model.

Run history is short too: only the newest 20 records survive per routine, so a weekly review needs a copy of that history saved somewhere else.

That review is the job where a long-context model earns its place.

How to read a score before you trust a model with a job

Start with the average, then open the task list.

Each benchmark page shows every scored build, best first, with the judge's comment on hover.

Look at the worst rows as hard as the best ones, because an unattended job meets the worst case sooner or later.

Then check the head-to-head records against the model you're replacing.

Grok 4.7, for example, trails Hermes MoA 7-10 with 3 ties and trails Kimi K3 5-12 with 3 ties on its benchmark page.

If you want to see how members choose models for their own scheduled jobs, the AI Profit Boardroom has the trainings and the coaching calls for it.

Which plans carry Grok Bot

Grok Bot is part of Premium, Super and Ultra on The X Subscription; Plus doesn't include it.

For the trainings and the Agent OS to run model-pinned scheduled jobs, join the AI Profit Boardroom.

Also On Our Network

• Agent Operatorsthe six-step setup order from first manual run to tested routine
• AI Income Deskthree small-business routine setups and the usage cost of each run
• AI Tool Verdictan eight-criterion scorecard that rates routines 8 out of 10
• agentos.guidefive Agent OS prompts that plan, count and review every routine
• aiprofitboardroom.comthe complete Grok Bot Routines guide on the AI Profit Boardroom blog

Pick a top-board model to write and review the work, let Grok Bot run it, and pin your own model on any Hermes copy, and Grok Bot routines get the best of both; book a free SEO strategy session if you want help setting it up.